Probably has raised $9 million in seed funding from Andreessen Horowitz to develop AI systems that prevent hallucinations and factual errors from reaching users. Founder Peter Elias said the company is targeting 99.99% accuracy, a standard common in deterministic software but difficult to achieve with large language models.
The startup’s first product is an AI data science tool that answers questions about complex datasets. Each response includes citations and an audit trail showing how the system reached its result.
Validator System Checks AI-Generated Answers
Probably uses a deterministic validator to compare the AI model’s initial answer with the underlying dataset. When the result does not match the data, the validator rejects it and returns it to the model for correction.
The company has also trained the model to work with the validator. Probably describes the surrounding system as a “data science mech suit” that limits ambiguity and guides the model toward answers supported by the source data.
Elias said stronger supporting infrastructure allows Probably to use less capable models without reducing accuracy. The current product operates on a model he described as four classes below the most advanced models available.
Running smaller models also allows the software to operate on local hardware, including desktop computers, rather than relying entirely on data centres. This can reduce the token and computing costs associated with processing repeated AI requests.
“What we learned building this was that the better your harness engineering is, the weaker the model can be,” Elias said. “If you can refine the context enough, the model does not have to work very hard to do the right thing.”
Probably Targets Precision-Sensitive Work
The company initially focused on data analysis because users need numerical answers to match the information contained in their datasets. Probably’s platform is designed to work with formats ranging from individual CSV files to large data warehouses containing billions of rows.
Its approach could later be applied to accounting, healthcare, and other areas where small factual errors may have serious consequences. Elias described these as precision-sensitive use cases that require stronger verification than a standard AI-generated response.
Probably’s funding comes as businesses reconsider the cost of using large AI models at scale. The company previously reported that running its infrastructure through DigitalOcean reduced its hardware costs by about 25% compared with a larger cloud provider.
The $9 million seed round was led by Andreessen Horowitz, according to TechCrunch. Probably has not disclosed the valuation attached to the financing or provided details about other participating investors.
Featured image credits: Magnific.com
For more stories like it, click the +Follow button at the top of this page to follow us.